yc-startup-rag-chatbot

🚀 YC Startup RAG Chatbot

A Retrieval-Augmented Generation (RAG) chatbot trained on Y Combinator’s “How to Start a Startup” lectures.

This project is an end-to-end RAG system that allows users to ask questions about YC’s Startup School lectures and get accurate, grounded answers with citations from the original transcript. It uses:

Python

Ollama (local LLM inference)

BGE-M3 embeddings

Streamlit (interactive chat UI)

Custom chunking + vector search

Local Retrieval-Augmented Generation pipeline

🧠 Features

✅ Chunking & Embeddings

Lecture transcripts are chunked using a Recursive Character Text Splitter.

Embeddings generated using BGE-M3 via Ollama.

Stored efficiently in embeddings.joblib.

✅ RAG Pipeline

Retrieve top-K most relevant chunks using cosine similarity.

Construct structured prompts using retrieved YC lecture content.

Generate grounded answers using lightweight local LLMs:

llama3.2:1b

✅ Interactive Chat UI

Built with Streamlit, showing:

Chat messages

Retrieved lecture chunks (sources)

Clean and readable answers

Session chat history

✅ Local, Privacy-Friendly & Fast

Everything runs fully offline using Ollama on your machine.

📂 Project Structure

├── app.py # Streamlit chat application ├── chunking.py # Splits transcripts into chunks ├── read_chunk.py # Generates embeddings + saves joblib ├── process_incoming.py # CLI-based RAG pipeline ├── requirements.txt ├── .gitignore ├── transcript/ # raw lecture transcripts (ignored) ├── json/ # chunk JSON files (ignored) ├── embeddings.joblib # embeddings (ignored) └── videos/, audios/ # raw data (ignored)

🚀 Getting Started

1️⃣ Clone the repository

git clone https://github.com//yc-startup-rag-chatbot.git cd yc-startup-rag-chatbot

2️⃣ Install dependencies

Create environment (optional):

conda create -n yc-rag python=3.10 -y conda activate yc-rag

Install packages:

pip install -r requirements.txt

3️⃣ Install and start Ollama

Download Ollama: https://ollama.com/download

Serve models:

ollama pull bge-m3 ollama pull llama3.2:1b # fastest

or ollama pull phi3 # best speed + quality

Start Ollama:

ollama serve

4️⃣ Prepare Data

Chunk transcripts

python chunking.py

Generate embeddings

python read_chunk.py

5️⃣ Run the Streamlit App

streamlit run app.py

Your browser will open the chatbot UI at:

http://localhost:8501

🧪 Example Questions

Try asking:

"What does Paul Graham say about generating startup ideas?"

"How should founders think about growth?"

"What is the most important quality in a co-founder?"

The app will show:

The answer

The exact lecture chunks used as context

🏗️ RAG Architecture

User Query ↓ Create Embedding (BGE-M3) ↓ Vector Search (Cosine Similarity) ↓ Retrieve Top-K Lecture Chunks ↓ Build Structured Prompt ↓ Local LLM (Llama3.2 / Phi3) ↓ Grounded Answer + Sources

🧩 Technologies Used

Python

Streamlit

Ollama

BGE-M3 embeddings

Numpy / Pandas

Scikit-Learn

Joblib

🧑‍💻 Author

Ashwani Jha RAG Developer | Machine Learning | LLMs

LinkedIn: www.linkedin.com/in/ashwani-jha-03ab14311